GenAI Explainability in the Rejection Email Flow

Overview

At SEEK, I worked on an early-stage concept focused on one of the clearest candidate jobs to be done: getting personalised feedback after an application outcome. The opportunity sat at the intersection of candidate experience, platform engagement, and business outcomes. The challenge was to find a way to provide feedback that felt genuinely useful to candidates, while staying aligned with our broader ambition to make hiring simpler rather than more complex. This case study covers how I framed the problem, explored solution directions, built alignment in a sensitive space, and helped get an AI explainability proof of concept prioritised and launched.

The problem

A recurring need in the Rejection email flow experience was candidate closure. Through the candidate jobs to be done, one of the strongest unmet needs was clear: candidates wanted personalised feedback that helped them understand what happened and what to do next.

The core question became:

How might we provide feedback that is useful to the candidate so they can feel closure and stay motivated in the application journey?

This was important for two reasons.

From a candidate perspective, silence or generic rejection messaging can feel discouraging and create a dead end in the application journey.
From a business perspective, better feedback had the potential to improve candidate trust, ongoing engagement, and future action on platform.

From the outset, I wanted to find an approach that served both sides of the equation:

meaningful value for candidates
a solution that did not introduce unnecessary friction for hirers

Discovery

I explored a few different ways we could address the problem.

1. Surface existing data

The first path was to understand whether there was already enough data in the system that we could repurpose into feedback for candidates. This was appealing because it could potentially unlock value without introducing new behaviour change.

The key question was whether the data we already had was specific and useful enough to feel personalised, rather than generic or obvious.

2. Add additional input steps on the hirer side

The second path was to ask hirers to provide more structured feedback that could then be passed back to candidates. On paper, this could create higher quality feedback. In practice, it raised a major concern: hirer intent to do this was low, and it would add another step into an experience where our core ambition was to make hiring simpler.

This quickly became one of the biggest trade-offs in the discovery phase. A solution might produce richer feedback, but if it relied on new hirer effort, it was unlikely to scale or align with product strategy.

3. Use AI explainability to support candidate outcomes

The third path was to explore whether AI-generated explanations could provide candidates with more meaningful feedback using data already available in the system. This approach was promising because it offered a way to deliver personalised guidance without requiring additional hirer action.

The upside was clear:

it could support the candidate JTBD directly
it stayed closer to the goal of reducing complexity for hirers
it created room to test whether explanation-style feedback would change candidate behaviour

After comparing the options, the two non-AI approaches were less compelling. Surfacing existing data alone risked being too shallow, and adding hirer steps was misaligned with both hirer intent and the strategic direction of simplifying hiring.

That made AI explainability the strongest concept to test.

Prioritisation

Shaping the hypothesis

Once we landed on AI explainability as the most viable path, I focused on where and how to test it. We decided to test the concept in a high-intent email. The reasoning was that candidates opening these emails were already in a moment where they were likely to pay attention and take action. If we could provide explanations that helped them understand their outcome and guide their next step, adoption was more likely than in a lower-intent surface.

The hypothesis was:

If we surface AI-generated explanations in a high-intent email, candidates will be more likely to engage with the content because it helps them understand what happened and what to do next. This should lead to higher open and click behaviour, and potentially stronger downstream platform outcomes.

User testing and concept validation

To validate the direction, we conducted user testing comparing non-AI variants with AI variants. The goal was not just to test preference, but to understand what kind of feedback candidates actually found helpful.

A few themes came through clearly:

most candidates saw AI-generated insights as helpful
the most valuable content was focused on next steps
suggested improvements were more compelling than abstract explanations alone

This helped sharpen the concept from “explaining an outcome” to “explaining an outcome in a way that gives the candidate a constructive next move”.

Building buy-in

This was a sensitive product area, and that made prioritisation challenging. Because the experience touched candidate trust and used AI in a visible way, there was a higher bar for confidence and caution. It was not enough to have an interesting idea; I needed to show why this problem mattered, why the proposed path was the right trade-off, and how we could test it responsibly.

To get traction, I presented the problem space to leadership and partner teams, grounding the case in:

candidate jobs to be done
candidate outcomes and experience quality
business outcomes tied to engagement and continued platform use
the strategic fit of solving this without adding hirer friction

A big part of the work was turning the concept from “interesting but risky” into “worth testing because the opportunity is meaningful and the scope is controlled”.

I also consulted teams who had run similar experiments before, as well as partner teams who could help clarify:

what data was available
what constraints existed
which fields were reliable enough to use
what risks we needed to manage upfront

That cross-functional discovery helped move the conversation from abstract possibility to practical feasibility.

In the end, I was able to get the proof of concept prioritised.

Launch

Because this was a proof of concept, I wanted to keep the scope tight. The objective was to prove or disprove the hypothesis without introducing unnecessary technical complexity or budget.

That led to a few deliberate decisions.

Keep the prompt and data inputs streamlined

Rather than trying to incorporate every possible signal, we focused on the must-have data points. This kept the prompt design simpler, reduced implementation complexity, and made it easier to assess whether the core idea itself had value.

Start with a narrower audience

We limited the proof of concept to lower-fit candidates. The hypothesis here was that the impact would be greater for this group, both in terms of candidate engagement and potential business outcomes. If the explanations could help candidates better understand the gap and what to improve next, the feature would be more likely to create a noticeable difference.

Deliver with the right partners

Working closely with AI teams and partner teams, we were able to bring the proof of concept to life. My role was to ensure the scope stayed anchored to the original hypothesis, the candidate value remained clear, and the experiment was realistic to deliver.

Results

The proof of concept launched successfully and gave us a way to test whether AI explainability could improve the candidate experience in a meaningful, scalable way.

Key outcomes included:

successful cross-functional alignment to launch the experiment
validation from research that candidates found AI insights helpful, especially when focused on next steps and suggested improvements
a scoped implementation that balanced learning value with technical and budget constraints

Reflection

What I'm most proud of in this work was helping move the idea from an uncertain problem space into something concrete enough to prioritise and launch. The challenge was not only product design or experimentation. It was also about building conviction in a cautious environment, aligning teams around a meaningful candidate problem, and finding a testable path that was both strategically sound and operationally realistic.

For me, this was a strong example of product management at its best: identifying a real user need, making thoughtful trade-offs, and turning an ambiguous opportunity into a focused experiment that could drive both user and business learning.

← Back to projects